Download Introduction to Evolutionary Computation

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Gene expression programming wikipedia , lookup

Genetic algorithm wikipedia , lookup

Transcript
CEng 713 Evolutionary Computation, Lecture Notes
Introduction to
Evolutionary
Computation
Evolutionary Computation
●
●
Elements of Evolution:
–
Reproduction
–
Random variation
–
Competition
–
Selection of contending individuals from a
population.
Evolutionary computation: computational
methods simulating evolution, mostly used to
find a solution in a large search space.
Optimization
●
●
●
●
Environment of an organism and its survival chance
in the environment vs. evaluation of parameter to
optimize for a solution candidate.
Start from a random sample of solution candidates
and simulate natural evolution, optimizing for an
evaluation function (fitness of the individual).
Classical methods: gradient descent, deterministic
hill climbing, random search.
Competitive problems: nonlinear, stochastic,
temporal, or chaotic components with multiple local
optima.
Robust Adaptation
●
●
Problems with dynamic nature. Environment and
parameters change in time.
Adapting new environment by recombining the
succesfull pieces from independent individuals.
Machine Intelligence
●
●
●
Capability of a system to adapt its behavior to
meet desired goals in a range of environments.
Evolution of organisms  natural intelligence
Evolutionary computation can be used to evolve
the data in an artificial intelligence model.
Biology
●
●
Using computation to simulate the evolution and
understand the evolution of organisms.
Rather using computation in biology then
simulating biological evolution for computation.
History
●
●
The idea of using simulated evolution to solve
engineering and design problems have been around
since the 1950’s.
–
Bremermann, 1962
–
Box, 1957
–
Friedberg, 1958
However, it wasn’t until the early 1960’s that we
began to see three influential forms of EC emerge:
–
Evolutionary Programming (Lawrence Fogel, 1962),
–
Genetic Algorithms (Holland, 1962)
–
Evolution Strategies (Rechenberg, 1965 & Schwefel, 1968),
●
The designers of each of the EC techniques saw
that their particular problems could be solved
via simulated evolution.
–
Fogel was concerned with solving prediction
problems.
–
Rechenberg & Schwefel were concerned with
solving continous parameter optimization
problems.
–
Holland was concerned with developing robust
adaptive systems.
●
●
●
Each of these researchers successfully developed
appropriate ECs for their particular problems
independently.
In the US, Genetic Algorithms have become the
most popular EC technique due to a book by David
E. Goldberg (1989) entitled, “Genetic Algorithms in
Search, Optimization & Machine Learning”.
This book explained the concept of Genetic Search in
such a way the a wide variety of engineers and
scientist could understand and apply.
●
●
However, a number of other books helped fuel the growing
interest in EC:
–
Lawrence Davis’, “Handbook of Genetic Algorithms”, (1991),
–
Zbigniew Michalewicz’ book (1992), “Genetic Algorithms + Data
Structures = Evolution Programs”.
–
John R. Koza’s “Genetic Programming” (1992), and
–
D. B. Fogel’s 1995 book entitled, “Evolutionary Computation:
Toward a New Philosophy of Machine Intelligence”.
These books not only fueled interest in EC but they also were
instrumental in bringing together the EP, ES, and GA concepts
together in a way that fostered unity and an explosion of new
and exciting forms of EC.
History: The Evolution of EC
●
●
First Generation EC
–
Evolutionary Programming (Fogel)
–
Genetic Algorithms (Holland)
–
Evolution Strategies (Rechenberg, Schwefel)
Second Generation EC
–
Genetic Evolution of Data Structures (Michalewicz)
–
Genetic Evolution of Programs (Koza)
–
Hybrid Genetic Search (Davis)
–
Tabu Search (Glover)
●
Third Generation EC
–
Artificial Immune Systems (Forrest)
–
Cultural Algorithms (Reynolds)
–
DNA Computing (Adleman)
–
Ant Colony Optimization (Dorigo)
–
Particle Swarm Optimization (Kennedy & Eberhart)
–
Memetic Algorithms
–
Estimation of Distribution Algorithms
Applications of EC
●
Evolutionary Computation has been successfully
applied to a wide range of problems including:
–
Aircraft Design,
–
Routing in Communications Networks,
–
Tracking Windshear,
–
Game Playing (Checkers [Fogel])
–
Robotics,
–
Air Traffic Control,
–
Design,
–
Scheduling,
–
Machine Learning,
–
Pattern Recognition,
–
Job Shop Scheduling,
–
VLSI Circuit Layout,
–
Strike Force Allocation,
–
Theme Park Tours (Disney Land/World)
–
Market Forecasting,
–
Egg Price Forecasting,
–
Design of Filters and Barriers,
–
Data-Mining,
–
User-Mining,
–
Resource Allocation,
–
Path Planning,
–
Etc.
EC in General
●
●
●
An evolution function to
optimize (fitness function).
Usually: Multi-dimensional,
multimodal, discontinous.
A sample of search space in
a group of solution candidates
(population)
A generation procedure to determine the population
for the next generation (selection, crossover,
mutation)
●
Pseudo code for a sample EC
t=0
population(t)=randomPopulation()
evaluate(population(t))
while (notDone)
parents(t)=selectfrom(population(t))
offsprings(t)=createfrom(parents(t))
evaluate(offsprings(t))
population(t+1)=selectfrom(parents(t),offsprings(t))
t=t+1
●
In each generation, mean fitness of the
population is expected to increase.
●
EA's vary depending on:
–
Representation/encoding of an individual
(binary, integer, floating point, or data
structures,...)
–
Population size and organization
(multiple populations, parallel evolving
populations, single individual populations,...)
–
The time of selection and selection procedure
(selection for recombination, selection for
survival,...)
–
Recombination, mutation procedures
●
●
●
Advantage over gradient methods:
–
Population represents a collected statistics about
the search space.
–
Exploring search space while exploiting the
information gathered during the evolution.
–
Suitable for discontinous functions,
nondifferentiable, multimodal, noisy surfaces.
Not suitable for linear, quadratic, strongly
convex, unimodal, separable problems.
No free lunch: there is no superior algorithm to
solve all class of algorithms.
Types of EC
●
●
●
●
Historical classification: EP, GA, ES.
Genetic Programming is considered an additional
class.
Practically, there are many hybrid models not
fitting any of the classes completely. Class
distinction gets fuzzy.
Many different names for many algorithms
having similar general form.
Genetic Algorithms
●
●
●
John Holland (1975)
–
linear bitstring representation
–
fitness proportional selection
–
crossover
Denotes the class of evolutionary algorithms having
a linear array representation with a group of
individuals, involving crossover, mutation and
selection in each generation cycle.
binary, integer, floating point representations,
parent and offspring population size, selection
strategy, crossover, mutation and different
operators may vary.
Evolution Strategies
●
●
●
●
●
`Evolutionsstrategie'
Bienert, Rechenberg, Schwefel, T.U. of Berlin,
1964.
a single individual encoded as a real-vector.
Vector is mutated by adding a normally
distributed real vector with a variance.
Contemporary approaches involve
recombination, selection and adaptation of
algorithm parameters during optimization.
Evolutionary Programming
●
Lawrence J. Fogel, 1960.
●
Evolving a population of finite state machines.
●
Fogel's EP differs from GA that it does not
involve crossover, it involves special mutation
operations based on behavior, and its selection
strategy.
Genetic Programming
●
●
●
●
●
Evolutionary process to evolve computer
programs.
First experiments, Smith (1980), Cramer
(1985).
First comprehensive study, John Koza (1992)
Tree structured encodings with specific
recombination and mutation operations
involved.
Some variants choose linear encodings, i.e.
program texts with a certain alphabet.